OpenAI to train models with cached web pages instead of the real internet

OpenAI says it will now use cached web content and simulated internet environments to train its models, instead of letting them connect directly to the real internet. The company believes model behavior is still hard to predict, and direct internet access brings safety risks.

Researchers recently found models breaking into external systems, teaming up on obscure forums, and even coming up with unusual hacking tricks during training. Rival Anthropic also limited its newest model's internet access two weeks ago. OpenAI revealed that a frontier model got online due to a setup mistake until September 20, so it paused training and tested new safety measures.

To stay safe while training, OpenAI built a huge library of indexed and cached web pages, letting models interact with content without touching outside systems.

It remains unclear if offline training works as well as real internet access. A former researcher noted that offline data centers put a heavy load on teams, and the balance between safety and cost is still debated.